See Spot Run

When I started Tempus, I knew that technology would disrupt healthcare. I had enough exposure to the impact technology had on other industries to see that writing on the wall. But unlike other sectors, healthcare had its own unique challenge that made the adoption of technology hard – data was siloed.  

So at the time, we began laying the foundation by amassing large multimodal datasets that included longitudinal data. We started in oncology and then migrated into other disease areas like cardiology, neurology, radiology, pathology, rare disorders and now immunology.    

Back then, “big data” was the buzzword of the day, and we were using conventional tools like natural language processing and optical character recognition to feed our machine learning efforts. I had no idea large language models were around the corner.   

A decade later, we’ve amassed one of the largest and most comprehensive multimodal datasets in all of healthcare as we approach 500 petabytes of connected data. We’ve been licensing this data for years to pharmaceutical companies and biotechs (all in ~ $2 billion over the past few years, given the valuable and scarce nature of this type of de-identified data).  

We’re super proud of our data business, including the impact it has had on drug discovery and development. And we’re proud that the world’s top pharma companies, like Merck, are signing up to become strategic partners joining the likes of AZ, GSK, BMS, etc. 

But the big technology companies, and the big frontier modeling firms, have been conspicuously absent. That was until recently. 

Lately, our phone has been “ringing” and as we peel through the data requests, everyone is asking the same question: “how much of your multimodal data has longitudinal patient journeys with outcome and response”

At first glance you might think this is because they want to ensure the completeness of data, but it’s not about quality or accuracy, it’s about prediction.

Just like large language models have been trained to predict the word “run” when someone types in “see spot” because it’s the most common word to follow, so too can models now be trained to predict what will come next for a patient if they have enough data to train. 

This is a holy grail, maybe the holy grail, for these models. 

Charging somebody $20 a month to rewrite an email is one thing. Charging $20 a month so a patient can avoid suffering because they are on the wrong drug is another. The first is susceptible to attrition, the second isn’t. 

The recent flurry of conversations we’re having is indicative of a chasm we have crossed.  These models are now sophisticated enough, and at sufficient scale, that if we feed them enough de-identified healthcare data (namely phenotypic, morphologic, and molecular healthcare data) at sequential time points with outcome and response, the models can predict what will come next. Pretty well at first, probably perfectly at some point.  

Here is a use case

“Hi X (pick your favorite AI App). I would like to provide you with a summary of my medical records which I just uploaded. As you can see, I’m a 56 year old man who has slightly elevated AIC. During my last three physicals over the past 3 years, my scores were 5.9, 5.8, and now 6.1. I’ve also recently increased the dose of the statin I’m on as I was taking 5mg of crestor, now I’m taking 10. My LDL is 121, and at my last appointment a year ago it was 103. Should I go on metformin as a precaution so I don’t become diabetic. What adverse events might I suspect? Any thoughts?”

Here is a likely response today:

Based on the A1C levels you shared (5.9, 5.8, and 6.1), you are currently in the prediabetes range, which is defined as 5.7% to 6.4%. An A1C of 6.5% or higher on two separate tests is generally the threshold for a diabetes diagnosis.

Regarding your question about metformin and your current statin use, here are some points to consider and discuss with your doctor:

Metformin is sometimes prescribed “off-label” to prevent the progression from prediabetes to Type 2 diabetes.

• Who benefits most? Clinical guidelines often suggest considering metformin for those at “high risk,” which typically includes people under 60 years old, those with a BMI ≥ 35, or those with an A1C ≥ 6.0%.
• Effectiveness: In major studies like the Diabetes Prevention Program (DPP), metformin reduced the risk of developing diabetes by about 31%, though intensive lifestyle changes (diet and exercise) were even more effective, reducing risk by 58%.
• Precautions: While generally safe, some experts suggest caution, noting that about one-third of people with prediabetes return to normal levels without medication, and two-thirds do not progress to diabetes for many years.

It is well-documented that statins, including Crestor (rosuvastatin), can cause a modest increase in blood sugar and A1C levels.

• Dose Dependency: The risk of elevated blood sugar is often greater with higher doses or high-intensity statin therapy.
• The Trade-off: Medical experts emphasize that the cardiovascular benefits of statins—preventing heart attacks and strokes—greatly outweigh the relatively small risk of a slight A1C increase. For example, one study found that while statins might lead to one extra case of diabetes for every 255 people treated, they prevented over five major cardiovascular events in that same group.

Suggested Next Steps

1. Shared Decision Making: Ask your doctor if your recent A1C uptick to 6.1 is likely related to the increased Crestor dose or if they recommend starting metformin now versus doubling down on lifestyle adjustments.
2. Lifestyle First: Since lifestyle modifications have shown higher efficacy than metformin in prevention trials, you might ask for a specific target for weight loss (often 5–7% of body weight) or physical activity (150 minutes/week).
3. Upload Records: I’m happy to review your medical records if you’d like to upload them! This would allow me to look for other risk factors like your BMI, triglycerides, or blood pressure that might influence a doctor’s decision to prescribe metformin.

While that is a nice response, it is far from personalized. Targeted, semi-intelligent – yes. Truly personalized, no.  

Here is the likely response in the future:

“Hi John. Thanks for connecting me to all of your medical records allowing me permission to review your medical history. Now that I have, you’re actually at moderate to high risk of being diagnosed with Type 2 Diabetes over the next several years. Your AIC is rising, but more importantly when you consider that your grandmother was a Type-2 diabetic, that increased risk factor along with your rising scores puts you in a higher risk category. That said, I’m worried about Metformin as an option for you given your age. You could experience muscle mass degradation and other side effects. I noticed based on your last physician encounter that you are still snacking regularly throughout the day and still eating candy with some regularity, so I would suggest you focus on diet first. Maybe talk to your doctor about ordering a glucose monitoring patch to see which foods are spiking your sugar levels. Finally, I would check your LDL score in the near future and see if the 10mg of Crestor brings down your LDL levels. In reading your charts, I noticed that you had another heart calcium scan recently and went from 0 calcification to 78. While that’s not alarmingly high, it does elevate your cardiac risk and so oIdeally your LDL levels are closer to 50 versus 121.”  

This future is coming. One way or another, it’s coming. I believe the big foundation model players can see that more clearly than the rest of us.  

To make that promise a reality, they need data, which companies like Tempus are in the best position to provide.

Otherwise, when applied to healthcare, See Spot Run might be See Spot Fish